A Comparative Evaluation of Machine Learning Models for an Enhanced Fraud Detection in Financial Systems

The increasing digitisation of financial transactions has led to a surge in fraudulent activities, posing significant challenges to the security, trust, and efficiency of financial systems. Existing rule-based systems struggle to adapt to evolving fraud tactics, which leads to high false positives and false negatives. Machine learning (ML) models offer a promising alternative, but limitations such as class imbalance, scalability, and interpretability hinder their practical deployment. This study evaluates and compares three ML models—Adaptive Boosting (AdaBoost), Decision Trees, and Logistic Regression—for fraud detection, focusing on their interpretability, scalability, and ability to handle class imbalance. A dataset of 284,807 anonymised financial transactions with a severe class imbalance was processed using SMOTE-Tomek for balancing and PCA for dimensionality reduction. The models were trained and optimised using grid search cross-validation, and their performance was evaluated using metrics such as accuracy, precision, recall, and F1-score derived from confusion matrices. Adaptive Boosting achieved the highest performance metrics, including 97.4% accuracy, 93% precision, recall, and F1-score, demonstrating its effectiveness in minimising false positives and false negatives. Decision Trees provided competitive accuracy (94.8%) while maintaining interpretability, and Logistic Regression, though computationally efficient, struggled with the imbalanced dataset, achieving an accuracy of 91.5%. The findings validate the efficacy of Adaptive Boosting for large-scale fraud detection systems and highlight the trade-offs between interpretability and performance in model selection. Recommendations for future research include integrating explainable AI techniques with high-performing models and extending evaluation to dynamic fraud datasets. Keywords: Adaptive Boosting, Class Imbalance, Fraud Detection, Interpretability, Machine Learning. CISDI Journal Reference Format Fatimah Adamu-Fika, Abdulrahman Suleiman Fatika, Aanuoluwapo Enyojo Baba-Onoja, Usman Adedayo Adeniyi, Henry Onyeoma Mafua, Onyinye Vivian Okpoko, Aisha Tijjani Ramalan (2025): A Comparative Evaluation of Machine Learning Models for an Enhanced Fraud Detection in Financial Systems. Computing, Information Systems, Development Informatics & Allied Research Journal. Vol 16 No 2, Pp 13-22. Available online at www.isteams.net/cisdijournal dx.doi.org/10.22624/AIMS/CISDI/V16N2P2